You can deploy this model and start earning money today!

GabeuxDev/
dactyl-ai-text-detector-onnx

Name: dactyl-ai-text-detector-onnx
Availability: InStock

onnx

deberta-v2

ai-generated-text-detection

text-classification

2508.00619

ShantanuT01/dactyl-ai-text-detector

quantized

mit

text-embeddings-inference

endpoints_compatible

Views

No views yet

DACTYL AI-Generated Text Detector, ONNX build

All credit for this model goes to the original authors. This repository is only a format conversion (PyTorch to ONNX) for CPU / onnxruntime inference. The weights and behaviour are unchanged from the original model.

Original model, please credit and cite

DACTYL, AI-Generated Text Detector, by Shantanu Thorat (@ShantanuT01)
Original model: https://huggingface.co/ShantanuT01/dactyl-ai-text-detector
Paper: https://arxiv.org/abs/2508.00619
Code: https://github.com/ShantanuT01/DACTYL
Base model: microsoft/deberta-v3-large
License: MIT (this conversion inherits the same license)

DACTYL is a DeBERTa-v3-large classifier fine-tuned with Empirical X-Risk Minimization (EXM) to optimise detection performance at low false-positive rates, with strong out-of-distribution generalisation to unseen models. Please see the original repository and paper for the full description, training data, and results.

What is in this repository

A faithful format conversion, nothing more:

onnx/model.onnx + onnx/model.onnx_data: the ONNX graph and weights (external-data format), exported with torch.onnx at opset 17.
tokenizer.json, tokenizer_config.json, config.json: copied unchanged from the original model.
turbochick.json: small runtime metadata (max sequence length).

The model outputs a single logit; apply a sigmoid to obtain the probability that the text is machine-generated. The conversion was validated to match the original PyTorch model to within 1e-5 on a set of sample texts.

Usage (onnxruntime)

python

import onnxruntime as ort
from tokenizers import Tokenizer

tok = Tokenizer.from_file("tokenizer.json")
sess = ort.InferenceSession("onnx/model.onnx", providers=["CPUExecutionProvider"])
enc = tok.encode("some text to score")
import numpy as np
ids = np.array([enc.ids], dtype=np.int64)
mask = np.array([enc.attention_mask], dtype=np.int64)
logit = sess.run(None, {"input_ids": ids, "attention_mask": mask})[0]
prob_machine = 1 / (1 + np.exp(-logit[0][0]))

License and attribution

MIT, inherited from the original DACTYL model. If you use this conversion, please cite the DACTYL paper (arXiv:2508.00619) and credit the original authors. This repo performs the ONNX conversion only.

GabeuxDev/dactyl-ai-text-detector-onnx